Thread

Bug in pg_stat_statements

Konstantin Knizhnik <knizhnik@garret.ru> — 2025-10-23T16:21:33Z

Hi hackers.


In PG18 gthere is bug either in pg_stat_statements.c, either in 
queryjumblefuncs.c.

Repro (you need Postgres build with asserts and pg_stat_statements 
included in shared_preload_librarties:

create function f(a integer[], out x integer, out y integer) returns 
record as $$            begin
        x = a[1];
        y = a[2];
    end;
$$ language plpgsql;

create table t(x integer);
select ((f(array[1,2]))).* from t;

The problems is caused by ((...)).* which cause to include the same 
array literal twice in JumbleQuery.
But pg_stat_stataement assumes that any occurrence of constant is unique 
and it cause assertion failure here:

         /* Copy next chunk (what precedes the next constant) */
         len_to_wrt = off - last_off;
         len_to_wrt -= last_tok_len;
         Assert(len_to_wrt >= 0);

So we should either exclude duplicates in RecordConstLocation 
(queryjumblefuncs.c)
either in generate_normalized_query in pg_stat_statements.c

What is considered to be more correct?